[models] optimize: call text_encoder.model instead of text_encoder to…

[xucqX]

In _encode_text of LTX2_T2AV_Adapter, calling self.pipeline.text_encoder(...) runs the full Gemma3ForConditionalGeneration.forward(), which unnecessarily computes logits through the LM head. Since we only need hidden_states, switching to self.pipeline.text_encoder.model(...) (i.e., Gemma3Model) skips the LM head projection, saving compute and memory while producing identical hidden states.

[Jayce-Ping]

I don't think this optimization is necessary. My reasons:

1. Maintaining encoder.__call__ is intentional to support future features that require a distributed backend wrapper. For example, when the text-encoder and the transformer are trained simultaneously.

2. The encode_prompt function is usually used in the preprocessing stage, where the computational overhead is quite small compared to training, calling encoder.model() directly will bypass the backend forward/backward hooks.